Aroma attribute intensity analysis method and device

Abstract

The invention relates to the field of aroma analysis, and provides an aroma attribute intensity analysis method and device in order to realize automatic analysis of aroma attribute intensity, direct conversion from instrument data to aroma attribute intensity is realized by constructing a quantitative correlation model of GC-IMS chromatographic information and aroma attribute intensity, manual auxiliary judgment is not needed, and the efficiency is high. Therefore, automatic analysis of aroma attribute intensity is realized; and based on automatic processing, the processing efficiency and consistency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of aroma analysis, in particular to an aroma attribute intensity analysis method and device. BACKGROUND

[0002] The aroma of baijiu is a core indicator of its quality, style and consumer acceptance. Traditional baijiu aroma evaluation mainly relies on sensory evaluation by professional tasters, who manually score the aroma, taste and style in combination. However, this approach has significant limitations. 1) Strong subjectivity: the evaluation results are greatly influenced by the experience, physiological state and environmental factors of the tasters, and the aroma intensity scores of the same baijiu by different tasters can differ by 10-20%, making consistency difficult to guarantee; 2) Low efficiency: single sensory evaluation requires the taster to perform multiple steps of "smelling aroma-tasting-aftertaste", which is difficult to meet the rapid analysis needs of large-scale samples (such as production line batch detection and new product research and development screening); 3) Difficulty in quantification: sensory evaluation results are mostly qualitative or semi-quantitative descriptions (such as "stronger cellar aroma"), which cannot achieve precise numerical expression of aroma attribute intensity, and are not conducive to the standardization and inheritance of baijiu aroma characteristics.

[0003] To solve the above problems, instrument analysis techniques such as gas chromatography-mass spectrometry (GC-MS) and high-performance liquid chromatography (HPLC) have been gradually introduced to detect volatile aroma components (such as esters, alcohols and acids) in baijiu. However, such techniques still have defects: on the one hand, GC-MS and other devices are complex to operate, have long detection periods (0.5-1 hour for single detection), and require professional personnel for component analysis; on the other hand, the "chemical component content" detected by the instrument is not directly linearly related to the "aroma attribute intensity perceived by the human body" - some components have low content but low aroma threshold (such as certain aldehydes), which greatly contribute to aroma, while some components have high content but high aroma threshold (such as certain alcohols), which have little effect on aroma perception, resulting in the fact that existing instrument analysis cannot directly output the aroma attribute intensity of baijiu, and still requires manual sensory evaluation to complete aroma evaluation. SUMMARY

[0004] To achieve automatic analysis of aroma attribute intensity, the present application provides an aroma attribute intensity analysis method and device.

[0005] The technical solution adopted by the present application to solve the above problems is as follows:

[0006] The aroma attribute intensity analysis method comprises:

[0007] Step 1, constructing a sample set, the sample set comprising GC-IMS chromatograms and aroma attribute intensity scores corresponding to different types of samples, the GC-IMS chromatogram being a two-dimensional color pixel graph encoded in RGB color;

[0008] Step 2, determining n identification regions of the GC-IMS chromatogram;

[0009] Step 3, constructing an aroma attribute intensity prediction model and training the model based on the sample set, the aroma attribute intensity prediction model taking the determined GC-IMS chromatogram identification region as input and taking the aroma attribute intensity score as output;

[0010] Step 4, obtaining the GC-IMS chromatogram of the sample to be tested, extracting the identification region and obtaining the aroma attribute intensity score based on the aroma attribute intensity prediction model.

[0011] Further, the sample type includes: flavor type, brand and year.

[0012] Further, step 1 is specifically: using a GC-IMS analyzer to obtain a two-dimensional chromatogram of each liquor sample; using a multi-person parallel evaluation + mean correction method to obtain the aroma attribute intensity score of each sample.

[0013] Further, the identification region determination method includes expert annotation and automatic identification; the expert annotation refers to: a GC-IMS analysis expert annotates n key chromatographic region blocks as identification regions; the automatic identification refers to: performing signal intensity threshold screening on the GC-IMS chromatogram to obtain an effective signal region, drawing an external rectangle for each continuous effective region, and automatically obtaining n identification regions.

[0014] Further, the aroma attribute intensity prediction model includes a chromatogram feature reconstruction module and an aroma attribute prediction module,

[0015] The chromatogram feature reconstruction module takes the identification region as input and outputs the chromatogram reconstruction feature based on a convolutional neural network model;

[0016] The aroma attribute prediction module takes the chromatogram reconstruction feature as input and outputs the aroma attribute intensity score based on a CNN or MLP.

[0017] Further, before inputting the identification region into the chromatogram feature reconstruction module, it further includes: calculating the peripheral size of the identification region, constructing a square white background according to the maximum value in the peripheral size, and placing the n identification regions in the n white backgrounds to form n square RGB images; according to the preset input size of the chromatogram feature reconstruction module, resampling the n square RGB images, and taking the resampled images as the input of the chromatogram feature reconstruction module.

[0018] Further, it further includes drawing an aroma profile based on the aroma attribute intensity score obtained based on the aroma attribute intensity prediction model.

[0019] The aroma attribute intensity analysis device comprises:

[0020] The sample acquisition module is configured to construct a sample set, wherein the sample set comprises GC-IMS chromatograms corresponding to different types of samples and aroma attribute intensity scores, and the GC-IMS chromatogram is a two-dimensional color pixel map encoded in RGB color;

[0021] The recognition region determination module is configured to determine n recognition regions of the GC-IMS chromatogram.

[0022] The aroma attribute intensity prediction model module is configured to construct an aroma attribute intensity prediction model and train the model based on the sample set, wherein the aroma attribute intensity prediction model takes the determined GC-IMS chromatogram recognition region as input and takes the aroma attribute intensity score as output.

[0023] The prediction module is configured to acquire the GC-IMS chromatogram of a sample to be tested, extract the recognition region, and acquire the aroma attribute intensity score based on the aroma attribute intensity prediction model.

[0024] Further, the aroma attribute intensity prediction model comprises a chromatogram feature reconstruction module and an aroma attribute prediction module,

[0025] The chromatogram feature reconstruction module takes the recognition region as input and outputs chromatogram reconstruction features based on a convolutional neural network model.

[0026] The aroma attribute prediction module takes the chromatogram reconstruction features as input and outputs the aroma attribute intensity score based on a CNN or MLP.

[0027] Further, the system further comprises a profile drawing module configured to draw an aroma profile based on the aroma attribute intensity score acquired by the aroma attribute intensity prediction model.

[0028] The present application has the beneficial effects compared with the prior art: the present application constructs a quantitative correlation model between GC-IMS chromatographic information and aroma attribute intensity, realizes direct conversion from instrument data to aroma attribute intensity, and does not need manual assistance for judgment, thereby realizing automatic analysis of aroma attribute intensity; based on automatic processing, the processing efficiency and consistency are improved. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The aroma attribute intensity analysis method is shown in the flowchart. DETAILED DESCRIPTION

[0030] As a new separation and detection technology, gas chromatography-ion mobility spectrometry (GC-IMS) has high separation ability of gas chromatography and high sensitivity (detection limit can reach ppb level) of ion mobility spectrometry, and can directly capture the two-dimensional distribution information of volatile aroma components (with GC retention time as the vertical coordinate, IMS migration time as the horizontal coordinate, and signal strength represented by color), and is an ideal tool for aroma analysis. Based on this, the present application realizes the direct conversion from instrument data to aroma attribute intensity by constructing a quantitative correlation model of “GC-IMS chromatographic information” and “aroma attribute intensity”.

[0031] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0032] As shown in Figure 1 The aroma attribute intensity analysis method comprises:

[0033] Step 1, constructing a sample set, wherein the sample set comprises GC-IMS chromatograms and aroma attribute intensity scores corresponding to different types of samples.

[0034] Taking liquor aroma analysis as an example, a plurality of groups of different types of liquor samples are selected, including different types, different brands, different years, etc., and the number of samples in each group is ≥30, and the total sample amount is ≥300. A large number of samples are used to ensure the model generalization ability.

[0035] The GC-IMS chromatogram is a two-dimensional color pixel image encoded with RGB color. The GC-IMS chromatogram of each sample needs to keep the same IMS migration time range, the same GC retention time range, the same IMS migration time plotting interval, the same GC retention time plotting interval, the same color and numerical value representation standard.

[0036] In this embodiment, the full GC-IMS chromatogram of all samples keeps the same one-dimensional time range and the same two-dimensional time range; the same IMS migration time and GC retention time plotting interval; the same picture resolution is used to reflect the picture resolution, which is 1584 pixels x 809 pixels, and at the same time, the same color and numerical value representation standard is set when exporting the full two-dimensional gas chromatogram.

[0037] The aroma attribute intensity score of each sample was obtained by multi-person parallel evaluation + mean correction method: 12 wine tasting experts with wine tasting experience were trained to unify the scoring standard using reference samples, and each sample was independently scored according to a 1-10 point system (1 point = very weak, 10 points = very strong). Each sample was scored repeatedly 3 times, and the average value after removing outliers was taken as the true score label of the aroma attribute of the sample.

[0038] Step 2, determining the identification region of the GC-IMS chromatogram.

[0039] The determination method of the identification region includes expert annotation and automatic recognition. Expert annotation refers to: a GC-IMS analysis expert combines a liquor aroma component database to annotate n key chromatographic region blocks, each region block is a rectangle, and is located by the "GC retention time-IMS migration time" coordinates of two opposite corners. Automatic recognition refers to: the GC-IMS chromatogram is subjected to signal intensity threshold screening (such as setting the threshold value to 5% of the maximum signal intensity to filter background noise), and the effective signal region is obtained. An outer rectangle is drawn for each continuous effective region to automatically obtain n key region blocks. In automatic recognition, the key region blocks obtained from different GC-IMS chromatograms may not be the same. In order to ensure the integrity of the identification, the key region blocks obtained from multiple GC-IMS chromatograms can be combined as the final n key region blocks.

[0040] It should be noted that for the same batch analysis, the consistency of the n key region positions should be ensured. And the n key chromatographic region blocks obtained by the key region identification step need to ensure the same spatial resolution and cannot be compressed and deformed.

[0041] By determining the identification region, the background noise is effectively reduced, and the efficiency of the later model feature processing is improved.

[0042] Step 3, constructing an aroma attribute intensity prediction model and training the model based on the sample set, wherein the aroma attribute intensity prediction model takes the determined GC-IMS chromatogram identification region as input and takes the aroma attribute intensity score as output.

[0043] The aroma attribute intensity prediction model includes a chromatogram feature reconstruction module and an aroma attribute prediction module.

[0044] The chromatogram feature reconstruction module adopts a convolutional neural network model (CNN). The CNN model can be constructed using a pre-trained model (such as ResNet), or it can be built by itself. The chromatogram feature reconstruction module performs feature reconstruction operations on the n key region blocks. In the operation process, the chromatogram feature reconstruction module shares weight parameters. Or it can be considered that there are n shared weight parameters for the CNN model used for chromatogram feature reconstruction, which are used to perform feature reconstruction on the n key region blocks.

[0045] For each key region block, the input feature is the RGB image corresponding to that key region block, and the output feature is the chromatogram reconstruction feature.

[0046] Before being input into the chromatogram feature reconstruction module, the outer envelope of all n key region blocks is calculated. A square white background image is constructed based on the larger value of the outer envelope. The n key region blocks are then centered within these n square white background images, forming n square RGB images. Subsequently, the n square RGB images are resampled according to the preset input pixel size of the chromatogram feature reconstruction module. For example, this embodiment uses a pre-trained ResNet 18 network as the chromatogram feature reconstruction module, and its input feature pixel size is preset to a 224×224 RGB image.

[0047] By uniformly placing all key regions on a square white background and resampling them to a preset size, not only is the image size input to the chromatogram feature reconstruction module consistent, but the interference of background noise on feature extraction is also reduced, allowing the model to focus more on the features of the key regions themselves.

[0048] The output features of the chromatogram feature reconstruction module are the features abstracted by convolution operations. In this embodiment, the output features are taken as a 3×3×3 feature tensor.

[0049] The aroma attribute prediction module is built based on a CNN or a fully connected network (MLP). If it is based on a CNN, the n 3×3×3 feature tensors obtained in the previous step can be concatenated to construct a new N×N×3 concatenated tensor, where N is the set of features. Three times the integer obtained by rounding up. For non-integer tensors, any blank positions at the end should be filled with 0. If based on MLP, the n 3×3×3 feature tensors are flattened and concatenated into a one-dimensional tensor of length 27n. In this embodiment, an MLP is used to construct the aroma attribute prediction module.

[0050] During model training, the chromatogram feature reconstruction module and the aroma attribute prediction module are inseparable and trained together; and the data in the sample set are divided into a training set (for model weight learning) and a validation set (for model performance verification) in an 8:2 ratio.

[0051] Step 4: Obtain the GC-IMS chromatogram of the sample to be tested, extract the identification region, and obtain the aroma attribute intensity score based on the aroma attribute intensity prediction model.

[0052] Furthermore, it also includes drawing aroma profile maps, such as bar charts and radar charts, based on aroma attribute intensity scores obtained from aroma attribute intensity prediction models, to intuitively display the distribution of aroma characteristics; it also supports image export and report generation.

[0053] Correspondingly, the application also provides an aroma attribute intensity analysis device, comprising:

[0054] a sample acquisition module for constructing a sample set comprising GC-IMS chromatograms and aroma attribute intensity scores corresponding to different types of samples, the GC-IMS chromatogram being a two-dimensional color pixel map encoded in RGB color;

[0055] a recognition area determination module for determining n recognition areas of the GC-IMS chromatogram;

[0056] an aroma attribute intensity prediction model module for constructing an aroma attribute intensity prediction model and training the model based on the sample set, the aroma attribute intensity prediction model taking the determined GC-IMS chromatogram recognition areas as input and taking the aroma attribute intensity scores as output;

[0057] a prediction module for acquiring the GC-IMS chromatogram of a sample to be tested, extracting the recognition areas and obtaining the aroma attribute intensity scores based on the aroma attribute intensity prediction model.

[0058] Specifically, the aroma attribute intensity prediction model comprises a chromatogram feature reconstruction module and an aroma attribute prediction module,

[0059] the chromatogram feature reconstruction module taking the recognition areas as input and outputting chromatogram reconstruction features based on a convolutional neural network model;

[0060] the aroma attribute prediction module taking the chromatogram reconstruction features as input and outputting the aroma attribute intensity scores based on a CNN or MLP.

[0061] Further, it also comprises a profile drawing module for drawing an aroma profile based on the aroma attribute intensity scores obtained by the aroma attribute intensity prediction model.

Claims

1. A method of analyzing the intensity of aroma attributes, characterized by, The method comprises the following steps: Step 1, constructing a sample set comprising GC-IMS chromatograms and aroma attribute intensity scores corresponding to different types of samples, wherein the GC-IMS chromatogram is a two-dimensional color pixel map encoded in RGB color; Step 2, determining n identified regions of the GC-IMS chromatogram; Step 3, constructing an aroma attribute intensity prediction model and training the model based on the sample set, wherein the aroma attribute intensity prediction model takes the determined identified regions of the GC-IMS chromatogram as input and takes the aroma attribute intensity score as output; Step 4, obtaining the GC-IMS chromatogram of a sample to be tested, extracting the identified regions, and obtaining the aroma attribute intensity score based on the aroma attribute intensity prediction model.

2. The method of aroma attribute intensity analysis according to claim 1, characterized in that, The sample types include liquor type, brand, and year.

3. The method of aroma attribute intensity analysis according to claim 1, characterized in that, Step 1 specifically comprises: obtaining a two-dimensional chromatogram of each liquor sample by using a GC-IMS analyzer; and obtaining the aroma attribute intensity score of each sample by using a multi-person parallel evaluation and mean correction method.

4. The method of aroma attribute intensity analysis according to claim 1, characterized in that, The method for determining the identified regions comprises expert annotation and automatic identification; the expert annotation refers to that a GC-IMS analysis expert annotates n key chromatographic region blocks as identified regions; and the automatic identification refers to that effective signal regions are obtained by performing signal intensity threshold screening on the GC-IMS chromatogram, and n identified regions are automatically obtained by drawing an external rectangle for each continuous effective region.

5. The method of aroma attribute intensity analysis according to claim 1, characterized in that, The aroma attribute intensity prediction model comprises a chromatogram feature reconstruction module and an aroma attribute prediction module, The chromatogram feature reconstruction module takes the identified regions as input and outputs chromatogram reconstruction features based on a convolutional neural network model; The aroma attribute prediction module takes the chromatogram reconstruction features as input and outputs the aroma attribute intensity score based on a CNN or MLP.

6. The method of aroma attribute intensity analysis according to claim 5, characterized in that, Before the identified regions are input into the chromatogram feature reconstruction module, the following steps are further included: calculating the peripheral size of the identified regions, constructing a square white background map according to the maximum value in the peripheral size, placing the n identified regions in the n white background maps to form n square RGB images; and resampling the n square RGB images according to the preset input size of the chromatogram feature reconstruction module, and taking the resampled images as the input of the chromatogram feature reconstruction module.

7. The method of aroma attribute intensity analysis according to claim 1, characterized in that, The method further comprises drawing an aroma profile based on the aroma attribute intensity score obtained based on the aroma attribute intensity prediction model.

8. An aroma attribute intensity analysis device, characterized by, The method comprises the following steps: A sample acquisition module is configured to construct a sample set comprising GC-IMS chromatograms and aroma attribute intensity scores corresponding to different types of samples, wherein the GC-IMS chromatogram is a two-dimensional color pixel map encoded in RGB color; An identified region determination module is configured to determine n identified regions of the GC-IMS chromatogram; An aroma attribute intensity prediction model module is configured to construct an aroma attribute intensity prediction model and train the model based on the sample set, wherein the aroma attribute intensity prediction model takes the determined identified regions of the GC-IMS chromatogram as input and takes the aroma attribute intensity score as output; A prediction module is configured to obtain the GC-IMS chromatogram of a sample to be tested, extract the identified regions, and obtain the aroma attribute intensity score based on the aroma attribute intensity prediction model.

9. The aroma attribute intensity analysis device according to claim 8, characterized by The aroma attribute intensity prediction model comprises a chromatogram feature reconstruction module and an aroma attribute prediction module, The chromatogram feature reconstruction module takes the identified region as input and outputs chromatogram reconstruction features based on a convolutional neural network model; The aroma attribute prediction module takes the chromatogram reconstruction features as input and outputs aroma attribute intensity scores based on a CNN or MLP.

10. The aroma attribute intensity analysis device according to claim 8, characterized by The profile plot drawing module also includes an aroma attribute intensity score drawn based on the aroma attribute intensity score obtained by the aroma attribute intensity prediction model.

Images